LET YOUR AGENT FIND THE EDGE
The Rust backtesting engine built for AI agents. Your agent loads your data, builds and event-driven backtests strategies, stress-tests them out-of-sample, and trains reinforcement-learning agents — live, in front of you, in a native macOS app. Connect Claude, Cursor, or any MCP agent.
One app · your agent · the whole research loop
Traditional vs. Agentic Research
How RLXBT changes the game for quant developers. Let your AI agent handle the heavy lifting while you direct the strategy.
Traditional Quant Loop
- •
Code Friction: Hours writing custom Pandas loops, managing package dependencies, and formatting CSV timelines.
- •
Vectorized Shortcuts: Prone to look-ahead bias and unrealistic trade fills, concealing real-world execution slippage.
- •
Curve-Fitting Risk: Tedious to code Walk-Forward splits, leading many traders to deploy overfit models in live markets.
- •
Isolated Runs: Backtests end up as scattered CSVs or local logs, forgotten or repeated due to lack of visual history.
Agentic Research Loop
- •
AI-Driven Coding: Ask Cursor or Claude Desktop in plain English. The agent loads the data, compiles rules, and runs the tests.
- •
Event-Driven Engine: Pure Rust simulator executing 6.6M bars/sec with intrabar exits, realistic latency, and tick accuracy.
- •
Automated Rigor: One click (or tool call) runs out-of-sample Walk-Forward and Monte Carlo simulations to prove your edge.
- •
Idea Map Integration: Visual spatial canvas tracks plan lineage, notes, and results, keeping the agent from repeating failures.
The whole research loop — automated
Your agent doesn't just generate a strategy. It backtests it, proves it out-of-sample, learns from the market, and shows you what actually holds up.
Your agent drives it
ActiveConnect Claude, Cursor, or any MCP agent. 30+ tools let it load data, build strategies, backtest, and train RL autonomously.
Watch it live
Every step renders in a native app: strategies, equity curves, and training reward curves. You see the agent think — not a black box.
Proof, not curve-fitting
Walk-forward, Monte-Carlo, sensitivity, and multi-strategy portfolio analysis. A robustness verdict tells you what holds up out-of-sample.
Reinforcement learning, built in
Train DQN trading agents in pure Rust — watch the live reward curve climb, verify IS/OOS splits, and query the model for forecasts.
Curate hundreds of runs
Every backtest is auto-archived in SQLite. Pin the winners, compare parameters, and build multi-strategy weighted portfolios.
Rust engine underneath
6.6M bars/sec throughput, realistic execution models, intrabar precision, and 10GB+ mmap loading with zero environment hassle.
Fast enough to
explore everything
A native Rust engine runs a full event-driven simulation — intrabar exits, realistic fills — fast enough for your agent to sweep hundreds of strategies and train RL agents while you watch.
Full event-driven bar-by-bar simulation — no vectorized shortcuts.
You give the idea. The agent does the work.
Connect your agent over MCP and tell it what to research. It runs the full loop and surfaces results in the app — you stay in the loop, not in the weeds.
Build & backtest
The agent inspects your data, drafts a strategy, validates the rules, and runs the backtest — every run archived as a report.
Prove it's not overfit
Walk-forward, Monte-Carlo risk-of-ruin, parameter sensitivity and multi-strategy portfolio analysis — the full suite that tells you whether the edge is real or just curve-fit to the past.
Train an RL agent
Spin up a reinforcement-learning trader that learns from the market — watch the reward curve climb and check it out-of-sample.
Ask for the forecast
“What's the call right now?” The trained model runs on the latest window and returns a long / short / flat signal in real time.
The Spatial Canvas
Idea Map
Stop running blind backtests. The Idea Map connects hypotheses, plans, and validated reports on a single infinite canvas.
Shared Context for AI Agents
Traders and agents brainstorm, draw connections, and set direction together. The map becomes the immediate context for your MCP agent, showing exactly which paths have failed and which show promise.
Compare & Refine Reports
Every backtest runs as a report, linked visually. Promising results turn green, rejected setups turn red, creating a clear, graphical roadmap of your quant research history.
Human-Agent Hybrid Loop
You steer the strategy direction, and the agent does the heavy data sweeps, walk-forward testing, and DQN models. Together, you build ideas that hold up out-of-sample.
Connect your agent in two minutes
Open the app, point Claude Desktop or Cursor at its MCP endpoint, and start a conversation. No code, no notebooks — just tell the agent what to test.
- ✓ Works with any MCP agent (Claude, Cursor, …)
- ✓ 30+ tools — backtest, stress-test, optimize, train RL, predict
- ✓ Everything the agent does shows up live in the app
Add new MCP server in settings:
Stop guessing. Let your agent prove it.
Download the Mac app, connect your agent, and run your first researched, out-of-sample-validated strategy today.
Free Sandbox to start · Starter $19/mo · Pro $49/mo · macOS (Apple Silicon)